ResNet Alzheimer's disease classification method based on fusion attention block

Abstract

The invention discloses a ResNet Alzheimer's disease classification method based on a fusion attention block, and aims to improve the diagnosis accuracy of Alzheimer's disease by improving a deep learning network architecture. The method comprises the following steps: firstly, acquiring and preprocessing MRI image data of a patient with Alzheimer's disease, including skull removal, registration to an MRI standard space, image smoothing and gray normalization; then, a ResNet34 basic network model is improved, a convolution block attention module is added to enhance a feature extraction effect, and an Meta-Acon activation function is used to process abnormal values and noise data; then, inputting the preprocessed image data into the improved ResNet34 model, extracting features through a convolutional layer, a pooling layer and a full connection layer, and finally obtaining a classification result through SoftPool pooling and a Meta-Acon activation function; and finally, classifying the features by using an improved ResNet34 model, and judging the Alzheimer's disease course stage of the subject.

Description

Technical Field

[0001] The present invention relates to the fields of medical image processing and artificial intelligence technology, and in particular to a ResNet Alzheimer's disease classification method based on fused attention blocks. Background Art

[0002] Alzheimer's disease is a serious neurodegenerative disease characterized by progressive and irreversible loss of cognitive function, which mainly affects the elderly population. With the intensification of the global aging problem, the prevalence of AD has shown a significant upward trend, which has brought huge economic and mental burdens to patients and their families. At present, the diagnosis of AD mainly relies on clinical symptoms, cognitive tests and neuroimaging examinations. However, since the early symptoms of Alzheimer's disease are relatively hidden, many patients are already in the middle and late stages of the disease when diagnosed. Therefore, early diagnosis is particularly important.

[0003] Traditional AD diagnostic methods include structural magnetic resonance imaging (MRI), positron emission tomography (PET), and cerebrospinal fluid biomarker detection. Although these methods have improved the accuracy of diagnosis to a certain extent, they still have some obvious defects: First, the diagnosis of MRI and PET images is highly dependent on the experience and expertise of clinicians. Manual reading is time-consuming and easily affected by subjective factors, which may lead to misdiagnosis or missed diagnosis. Secondly, PET scanning and cerebrospinal fluid biomarker detection are not only expensive, but also invasive and uncomfortable, which limits their application in large-scale screening. Finally, traditional image processing methods have limitations in feature extraction, making it difficult to fully utilize the information in the image, resulting in low diagnostic accuracy.

[0004] In response to the above problems, computer-aided diagnosis methods based on deep learning have gradually become a research hotspot in recent years. Deep learning, especially convolutional neural networks, has shown great potential in medical image analysis. It can automatically learn features from large amounts of data and reach or even exceed the level of human experts in classification tasks. However, the existing deep learning methods still have the following shortcomings in the diagnosis of Alzheimer's disease: when extracting features, traditional convolutional neural networks tend to ignore local important information in the image, resulting in insufficient feature extraction and affecting classification performance; conventional maximum pooling or average pooling methods tend to lose important information when downsampling features, resulting in a decrease in feature representation ability and affecting the overall performance of the model; commonly used activation functions have limitations when processing outliers and noisy data, which can easily lead to gradient vanishing or gradient explosion problems, affecting the training effect and stability of the model. Therefore, how to provide a ResNet Alzheimer's disease classification method based on fused attention blocks is a problem that technicians in this field urgently need to solve. Summary of the invention

[0005] One purpose of the present invention is to propose a ResNet Alzheimer's disease classification method based on fused attention blocks, aiming to improve the diagnostic accuracy of Alzheimer's disease by improving the deep learning network architecture, effectively solving the problem of inaccurate diagnosis results in the prior art, and significantly improving the diagnostic accuracy of Alzheimer's disease.

[0006] According to an embodiment of the present invention, a ResNet Alzheimer's disease classification method based on fusion attention blocks includes the following steps:

[0007] S1. Obtain MRI image data of patients with Alzheimer's disease and preprocess the MRI image data, including skull removal, registration to MRI standard space, image smoothing and grayscale normalization;

[0008] S2. Improve the ResNet34 basic network model:

[0009] (1) The convolutional block attention module is introduced into the ResNet34 basic network model to extract features by combining the channel attention mechanism and the spatial attention mechanism;

[0010] (2) Using the SoftPool pooling layer to replace the traditional maximum pooling or average pooling layer, and obtaining features through the weighted pooling process;

[0011] (3) The original ReLU activation function is replaced by the Meta-Acon activation function, and outliers and noise data are processed through adaptive learning;

[0012] S3. Introduce the convolutional block attention module to process preliminary features by combining the channel attention mechanism and the spatial attention mechanism;

[0013] S4. Use the SoftPool pooling layer to perform weighted pooling on the extracted features, and obtain features through the weighted pooling process;

[0014] S5. Input the features into the Meta-Acon activation function to process outliers and noise data through adaptive learning;

[0015] S6. Use the improved ResNet34 model to classify the processed features and determine whether the subject has Alzheimer's disease and the stage of the disease;

[0016] S7. Output a diagnosis report based on the classification results, showing the classification results of Alzheimer's disease and the corresponding disease stages.

[0017] Optionally, the S1 includes the following steps:

[0018] S11. Obtain MRI imaging data of patients with Alzheimer's disease;

[0019] S12. performing skull removal processing on the acquired MRI image data to remove interference of the skull on the image data;

[0020] S13. registering the MRI image data after skull removal to the MNI standard space;

[0021] S14. Perform image smoothing on the registered MRI image data:

[0022]

[0023] Among them, G(x,y) represents the pixel value of the image after Gaussian smoothing, and σ represents the standard deviation of the Gaussian kernel.

[0024] S15. Perform grayscale normalization on the smoothed MRI image data to normalize the pixel values ​​to the range of [0,1]:

[0025]

[0026] Among them, I′ represents the normalized pixel value, I represents the original pixel value, and I min and I max Represent the minimum and maximum pixel values ​​of the image respectively.

[0027] Optionally, S3 includes the following steps:

[0028] S31. Input the initially extracted features into the convolutional block attention module, which processes the features through the channel attention mechanism and the spatial attention mechanism:

[0029]

[0030] Among them, M c (F) represents the channel attention weight, σ represents the Sigmoid function, and They represent the weights of the channel convolution operation, AvgPool(F) and MaxPool(F) represent the average pooling and maximum pooling operations of the input feature F respectively;

[0031] F′=M c (F)·F;

[0032] Among them, F′ represents the feature after channel attention processing, M c (F) represents the channel attention weight, and F represents the input feature;

[0033] M s (F′)=σ(f 7×7 ([AvgPool(F′);MaxPool(F′)]))

[0034] Among them, M s (F′) represents the spatial attention weight, σ represents the Sigmoid function, and f 7×7 represents a 7x7 convolution operation, AvgPool(F′) and MaxPool(F′) represent the features F after channel attention processing respectively. ′ Average pooling and maximum pooling operations;

[0035] S32. Use the spatial attention mechanism to weight the features after channel attention processing:

[0036] F″=M s (F′)·F′

[0037] Among them, F″ represents the feature after spatial attention processing, M s (F′) represents the spatial attention weight, and F′ represents the feature after channel attention processing.

[0038] Optionally, S4 includes the following steps:

[0039] S41. Input the features processed by the convolutional block attention module into the SoftPool pooling layer for weighted pooling operation:

[0040]

[0041] Among them, SoftPool(x) represents the feature value after pooling, x i represents the eigenvalue within the pooling window, and e represents the natural exponent;

[0042] S42. Calculate the weight of the feature value in each pooling window:

[0043]

[0044] Among them, w i represents the eigenvalue x i The weight of Represents the exponential sum of all eigenvalues ​​within the pooling window;

[0045] S43. Perform weighted summation on the eigenvalues ​​in the pooling window using the calculated weights to obtain the eigenvalues ​​after weighted pooling;

[0046] S44. Take the feature values ​​after weighted pooling as output for subsequent feature processing and classification.

[0047] Optionally, S5 includes the following steps:

[0048] S51. Input the weighted pooled features into the Meta-Acon activation function for activation processing;

[0049] β=σ(W1·GlobalAvgPool(F)+W2;

[0050] Among them, β represents the adaptive parameter, σ represents the Sigmoid function, W1 and W2 are the weight matrices of the 1x1 convolution layer, GlobalAvgPool represents the global average pooling operation, and F represents the input feature;

[0051] S52. Use the calculated adaptive parameter β to activate the weighted pooled features:

[0052] Meta-Acon(x)=x·σ(β·x);

[0053] Among them, σ represents the Sigmoid function, β represents the adaptive parameter, and x represents the input feature;

[0054] S53. The processed feature values ​​are used as the final output for subsequent classification operations.

[0055] Optionally, the S6 includes the following steps:

[0056] S61. Input the features processed by the Meta-Acon activation function into the fully connected layer of the improved ResNet34 model;

[0057] O = W·F + b;

[0058] Among them, O represents the output feature of the fully connected layer, W represents the weight matrix of the fully connected layer, F represents the input feature, and b represents the bias term;

[0059] S62. Normalize the output features of the fully connected layer and map them to the classification probability space;

[0060] S63. Use the Softmax function to classify the normalized features and calculate the probability value of each category:

[0061]

[0062] Among them, P(y=j|x) represents the probability that the input feature x belongs to category j, and W j represents the weight vector of category j, and K represents the total number of categories;

[0063] S64. judging whether the subject suffers from Alzheimer's disease and the stage of the disease according to the calculated probability value as a classification result;

[0064] S65. Store and output the classification results for generating a final diagnosis report.

[0065] The beneficial effects of the present invention are:

[0066] (1) Improve diagnostic accuracy: By introducing the convolutional block attention module (CBAM) into the basic network model, the present invention can more effectively extract key features from MRI images. Combining the channel attention mechanism and the spatial attention mechanism, the focus on important features is enhanced, thereby improving the classification accuracy of Alzheimer's disease.

[0067] (2) Reduce information loss: The SoftPool pooling layer is used to replace the traditional maximum pooling or average pooling layer, and more fine-grained features are obtained through the weighted pooling process. This method can retain more detailed information, reduce information loss, make the feature representation more comprehensive and accurate, and thus improve classification performance.

[0068] (3) Enhanced model robustness: The original ReLU activation function is replaced with the Meta-Acon activation function, and the robustness and stability of the model are enhanced by processing outliers and noisy data through adaptive learning. When processing complex and variable data, the model of the present invention can maintain high performance and reduce errors caused by data anomalies and noise.

[0069] (4) Early diagnosis advantage: The method of the present invention can provide more accurate diagnosis in the early stages of Alzheimer's disease through an improved deep learning network architecture. This is of great significance for the early intervention and treatment of the disease, which can delay the progression of the disease and improve the quality of life of patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0071] Figure 1 This is a flowchart of a ResNet Alzheimer's disease classification method based on fused attention blocks proposed in the present invention. DETAILED DESCRIPTION

[0072] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0073] refer to Figure 1 , a ResNet Alzheimer's disease classification method based on fusion attention blocks, including the following steps:

[0074] S1. Obtain MRI image data of patients with Alzheimer's disease and preprocess the MRI image data, including skull removal, registration to MRI standard space, image smoothing and grayscale normalization;

[0075] S2. Improve the ResNet34 basic network model:

[0076] (1) The convolutional block attention module is introduced into the ResNet34 basic network model to extract features by combining the channel attention mechanism and the spatial attention mechanism;

[0077] (2) Using the SoftPool pooling layer to replace the traditional maximum pooling or average pooling layer, and obtaining features through the weighted pooling process;

[0078] (3) The original ReLU activation function is replaced by the Meta-Acon activation function, and outliers and noise data are processed through adaptive learning;

[0079] S3. Introduce the convolutional block attention module to process preliminary features by combining the channel attention mechanism and the spatial attention mechanism;

[0080] S4. Use the SoftPool pooling layer to perform weighted pooling on the extracted features, and obtain features through the weighted pooling process;

[0081] S5. Input the features into the Meta-Acon activation function to process outliers and noise data through adaptive learning;

[0082] S6. Use the improved ResNet34 model to classify the processed features and determine whether the subject has Alzheimer's disease and the stage of the disease;

[0083] S7. Output a diagnosis report based on the classification results, showing the classification results of Alzheimer's disease and the corresponding disease stages.

[0084] In this implementation, S1 includes the following steps:

[0085] S11. Obtain MRI imaging data of patients with Alzheimer's disease;

[0086] S12. performing skull removal processing on the acquired MRI image data to remove interference of the skull on the image data;

[0087] S13. registering the MRI image data after skull removal to the MNI standard space;

[0088] S14. Perform image smoothing on the registered MRI image data:

[0089]

[0090] Among them, G(x,y) represents the pixel value of the image after Gaussian smoothing, and σ represents the standard deviation of the Gaussian kernel.

[0091] S15. Perform grayscale normalization on the smoothed MRI image data to normalize the pixel values ​​to the range of [0,1]:

[0092]

[0093] Among them, I′ represents the normalized pixel value, I represents the original pixel value, and I min and I max Represent the minimum and maximum pixel values ​​of the image respectively.

[0094] In this implementation, S3 includes the following steps:

[0095] S31. Input the initially extracted features into the convolutional block attention module, which processes the features through the channel attention mechanism and the spatial attention mechanism:

[0096]

[0097] Among them, M c (F) represents the channel attention weight, σ represents the Sigmoid function, and They represent the weights of the channel convolution operation, AvgPool(F) and MaxPool(F) represent the average pooling and maximum pooling operations of the input feature F respectively;

[0098] F′=M c (F)·F;

[0099] Among them, F′ represents the feature after channel attention processing, M c (F) represents the channel attention weight, and F represents the input feature;

[0100] M s (F′)=σ(f 7×7 ([AvgPool(F′);MaxPool(F′)]))

[0101] Among them, M s (F′) represents the spatial attention weight, σ represents the Sigmoid function, and f 7×7 represents a 7x7 convolution operation, AvgPool(F′) and MaxPool(F′) represent the average pooling and maximum pooling operations of feature F′ after channel attention processing, respectively;

[0102] S32. Use the spatial attention mechanism to weight the features after channel attention processing:

[0103] F″=M s (F′)·F′

[0104] Among them, F″ represents the feature after spatial attention processing, M s (F′) represents the spatial attention weight, and F′ represents the feature after channel attention processing.

[0105] In this implementation manner, S4 includes the following steps:

[0106] S41. Input the features processed by the convolutional block attention module into the SoftPool pooling layer for weighted pooling operation:

[0107]

[0108] Among them, SoftPool(x) represents the feature value after pooling, x i represents the eigenvalue within the pooling window, and e represents the natural exponent;

[0109] S42. Calculate the weight of the feature value in each pooling window:

[0110]

[0111] Among them, w i represents the eigenvalue x i The weight of Represents the exponential sum of all eigenvalues ​​within the pooling window;

[0112] S43. Perform weighted summation on the eigenvalues ​​in the pooling window using the calculated weights to obtain the eigenvalues ​​after weighted pooling;

[0113] S44. Take the weighted pooled feature values ​​as output for subsequent feature processing and classification.

[0114] In this implementation manner, S5 includes the following steps:

[0115] S51. Input the weighted pooled features into the Meta-Acon activation function for activation processing;

[0116] β=σ(W1·GlobalAvgPool(F)+W2;

[0117] Among them, β represents the adaptive parameter, σ represents the Sigmoid function, W1 and W2 are the weight matrices of the 1x1 convolution layer, GlobalAvgPool represents the global average pooling operation, and F represents the input feature;

[0118] S52. Use the calculated adaptive parameter β to activate the weighted pooled features:

[0119] Meta-Acon(x)=x·σ(β·x);

[0120] Among them, σ represents the Sigmoid function, β represents the adaptive parameter, and x represents the input feature;

[0121] S53. The processed feature values ​​are used as the final output for subsequent classification operations.

[0122] In this implementation manner, S6 includes the following steps:

[0123] S61. Input the features processed by the Meta-Acon activation function into the fully connected layer of the improved ResNet34 model;

[0124] O = W·F + b;

[0125] Among them, O represents the output feature of the fully connected layer, W represents the weight matrix of the fully connected layer, F represents the input feature, and b represents the bias term;

[0126] S62. Normalize the output features of the fully connected layer and map them to the classification probability space;

[0127] S63. Use the Softmax function to classify the normalized features and calculate the probability value of each category:

[0128]

[0129] Among them, P(y=j|x) represents the probability that the input feature x belongs to category j, and W j represents the weight vector of category j, and K represents the total number of categories;

[0130] S64. judging whether the subject suffers from Alzheimer's disease and the stage of the disease according to the calculated probability value as a classification result;

[0131] S65. Store and output the classification results for generating a final diagnosis report.

[0132] Embodiment 1:

[0133] In order to verify the feasibility of the present invention in implementation, the present invention was applied to the Department of Neurology of a large general hospital, which is mainly responsible for the screening and diagnosis of Alzheimer's disease in elderly patients. In this scenario, the hospital receives a large number of elderly patients every week, many of whom come for treatment due to memory loss or cognitive dysfunction. Traditional diagnostic methods mainly rely on the experience of doctors and manual reading of MRI images, but this method is time-consuming and easily affected by subjective factors, resulting in a high misdiagnosis rate. In order to improve this situation, the hospital decided to introduce the ResNet Alzheimer's disease classification method based on fused attention blocks, using deep learning technology to assist doctors in diagnosis.

[0134] Before applying the method of the present invention, it is first necessary to collect a large amount of MRI image data. These data come from the MRI image database of patients diagnosed with Alzheimer's disease, mild cognitive impairment and healthy elderly people in the hospital in the past five years. The data set includes MRI images of 3,000 patients, including 1,000 patients with Alzheimer's disease, 1,000 patients with mild cognitive impairment, and 1,000 healthy elderly people. The MRI images of each patient are strictly preprocessed, including skull removal, registration to MNI standard space, image smoothing and grayscale normalization.

[0135] After data preprocessing, the image data is input into the improved ResNet34 model for training. The model improvements include the introduction of convolutional block attention module, pooling layer and activation function. These improvements are aimed at improving the accuracy of feature extraction, reducing information loss, and enhancing the robustness of the model.

[0136] During the training process, the model gradually extracts the features of MRI images through multiple convolutional layers and pooling layers, and uses the CBAM module to further enhance the saliency of the features. The pooling layer retains more detailed information through the weighted pooling process, and the activation function handles outliers and noise data through adaptive learning. After dozens of rounds of iterative training, the model achieved a high classification accuracy.

[0137] After the model training was completed, we applied it to the actual diagnosis process of the hospital. In order to verify the effect of the model, a one-month period was selected for the experiment. During this period, the hospital received a total of 500 elderly patients, all of whom underwent MRI examinations and were diagnosed using the method of the present invention. In order to ensure the reliability of the results, all patients also received traditional diagnostic methods, and experienced doctors performed manual film reading. The accuracy of the method of the present invention is better than that of the manual film reading detection method. The specific data is shown in Table 1 below:

[0138] Table 1 Comparison between the manual film reading detection method and the method of the present invention

[0139]

[0140] The experimental results show that the method based on the present invention successfully diagnosed 320 Alzheimer's patients, 150 mild cognitive impairment patients, and 30 healthy elderly people among 500 patients. Compared with the traditional method, the accuracy of the method of the present invention in diagnosing Alzheimer's disease reached 95%, significantly higher than the 82% of the traditional method. In addition, the diagnostic accuracy of mild cognitive impairment was also improved from 70% of the traditional method to 88%. In the identification of healthy elderly people, the accuracy of the method of the present invention was 98%.

[0141] In order to further demonstrate the advantages of the method of the present invention in reducing manual errors and improving work efficiency, we counted the time doctors spent on manual film reading and diagnosis using the method of the present invention. The results showed that it took doctors an average of 15 minutes per patient to read the film manually, while it only took 5 minutes per patient to diagnose using the method of the present invention. This data shows that the method of the present invention not only greatly improves the diagnostic accuracy, but also significantly improves the diagnostic efficiency and reduces the workload of doctors.

[0142] The ResNet Alzheimer's disease classification method based on the fusion attention block of the present invention has shown significant advantages in practical applications. By improving the accuracy of feature extraction, reducing information loss, and enhancing the robustness of the model, the method of the present invention can more accurately perform early diagnosis of Alzheimer's disease. In addition, by greatly improving diagnostic efficiency and reducing manual errors, the method of the present invention not only improves the work efficiency of the hospital, but also provides patients with more timely and accurate diagnostic services. This innovative technical solution has broad prospects and important social significance in clinical applications.

[0143] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A ResNet Alzheimer's disease classification method based on fusion attention blocks, characterized in that: The following steps are involved: S1. Obtain MRI image data of patients with Alzheimer's disease and preprocess the MRI image data, including skull removal, registration to MRI standard space, image smoothing and grayscale normalization; S2. Improve the ResNet34 basic network model: (1) The convolutional block attention module is introduced into the ResNet34 basic network model to extract features by combining the channel attention mechanism and the spatial attention mechanism; (2) Using the SoftPool pooling layer to replace the traditional maximum pooling or average pooling layer, and obtaining features through the weighted pooling process; (3) The original ReLU activation function is replaced by the Meta-Acon activation function, and outliers and noise data are processed through adaptive learning; S3. Introduce the convolutional block attention module to process preliminary features by combining the channel attention mechanism and the spatial attention mechanism; S4. Use the SoftPool pooling layer to perform weighted pooling on the extracted features, and obtain features through the weighted pooling process; S5. Input the features into the Meta-Acon activation function to process outliers and noise data through adaptive learning; S6. Use the improved ResNet34 model to classify the processed features and determine whether the subject has Alzheimer's disease and the stage of the disease; S7. Output a diagnosis report based on the classification results, showing the classification results of Alzheimer's disease and the corresponding disease stages.

2. The Alzheimer's disease classification method based on ResNet fusion attention block according to claim 1, characterized in that: The S1 comprises the following steps: S11. Obtain MRI imaging data of patients with Alzheimer's disease; S12. performing skull removal processing on the acquired MRI image data to remove interference of the skull on the image data; S13. registering the MRI image data after skull removal to the MNI standard space; S14. Perform image smoothing on the registered MRI image data: Among them, G(x,y) represents the pixel value of the image after Gaussian smoothing, and σ represents the standard deviation of the Gaussian kernel. S15. Perform grayscale normalization on the smoothed MRI image data to normalize the pixel values ​​to the range of [0,1]: Among them, I′ represents the normalized pixel value, I represents the original pixel value, and I min and I max Represent the minimum and maximum pixel values ​​of the image respectively.

3. The Alzheimer's disease classification method based on ResNet fusion attention block according to claim 2 is characterized in that: The S3 comprises the following steps: S31. Input the initially extracted features into the convolutional block attention module, which processes the features through the channel attention mechanism and the spatial attention mechanism: Among them, M c (F) represents the channel attention weight, σ represents the Sigmoid function, and They represent the weights of the channel convolution operation, AvgPool(F) and MaxPool(F) represent the average pooling and maximum pooling operations of the input feature F respectively; F′=M c (F)·F; Among them, F′ represents the feature after channel attention processing, M c (F) represents the channel attention weight, and F represents the input feature; M s (F′)=σ(f 7×7 ([AvgPool(F′);MaxPool(F′)])) Among them, M s (F′) represents the spatial attention weight, σ represents the Sigmoid function, f 7×7 represents a 7x7 convolution operation, AvgPool(F′) and MaxPool(F′) represent the average pooling and maximum pooling operations of feature F′ after channel attention processing, respectively; S32. Use the spatial attention mechanism to weight the features after channel attention processing: F″=M s (F′)·F′ Among them, F″ represents the feature after spatial attention processing, M s (F′) represents the spatial attention weight, and F′ represents the feature after channel attention processing.

4. The Alzheimer's disease classification method based on ResNet fusion attention block according to claim 1, characterized in that: The S4 comprises the following steps: S41. Input the features processed by the convolutional block attention module into the SoftPool pooling layer for weighted pooling operation: Among them, SoftPool(x) represents the feature value after pooling, x i represents the eigenvalue within the pooling window, and e represents the natural exponent; S42. Calculate the weight of the feature value in each pooling window: Among them, w i Represents the eigenvalue x i The weight of Represents the exponential sum of all eigenvalues ​​within the pooling window; S43. Perform weighted summation on the eigenvalues ​​in the pooling window using the calculated weights to obtain the eigenvalues ​​after weighted pooling; S44. Take the weighted pooled feature values ​​as output for subsequent feature processing and classification.

5. The Alzheimer's disease classification method based on ResNet fusion attention block according to claim 1, characterized in that: The S5 comprises the following steps: S51. Input the weighted pooled features into the Meta-Acon activation function for activation processing; β=σ(W1·GlobalAvgPool(F)+W2; Among them, β represents the adaptive parameter, σ represents the Sigmoid function, W1 and W2 are the weight matrices of the 1x1 convolution layer, GlobalAvgPool represents the global average pooling operation, and F represents the input feature; S52. Use the calculated adaptive parameter β to activate the weighted pooled features: Meta-Acon(x)=x·σ(β·x); Among them, σ represents the Sigmoid function, β represents the adaptive parameter, and x represents the input feature; S53. The processed feature values ​​are used as the final output for subsequent classification operations.

6. The Alzheimer's disease classification method based on ResNet fusion attention block according to claim 1, characterized in that: The step S6 comprises the following sub-steps: S61. Input the features processed by the Meta-Acon activation function into the fully connected layer of the improved ResNet34 model; O = W·F + b; Among them, O represents the output feature of the fully connected layer, W represents the weight matrix of the fully connected layer, F represents the input feature, and b represents the bias term; S62. Normalize the output features of the fully connected layer and map them to the classification probability space; S63. Use the Softmax function to classify the normalized features and calculate the probability value of each category: Among them, P(y=j|x) represents the probability that the input feature x belongs to category j, and W j represents the weight vector of category j, and K represents the total number of categories; S64. judging whether the subject suffers from Alzheimer's disease and the stage of the disease according to the calculated probability value as a classification result; S65. Store and output the classification results for generating a final diagnosis report.

Images